On Multi-Dimensional Hilbert Indexings - Algorithmics and ...

LCNS, Vol 1449, pp. 329–338, Springer 1998

On Multi-Dimensional Hilbert Indexings Jochen Alber and Rolf Niedermeier? Wilhelm-Schickard-Institut fur Informatik, Universitat Tubingen, Sand 13, D-72076 Tubingen, Fed. Rep. of Germany, alber/[email protected]

Abstract. Indexing schemes for grids based on space- lling curves (e.g.,

Hilbert indexings) nd applications in numerous elds. Hilbert curves yield the most simple and popular scheme. We extend the concept of curves with Hilbert property to arbitrary dimensions and present rst results concerning their structural analysis that also simplify their applicability. As we show, Hilbert indexings can be completely described and analyzed by \generating elements of order 1", thus, in comparison with previous work, reducing their structural complexity decisively.

1 Introduction Discrete multi-dimensional spaces are of increasing importance. They appear in various settings such as combinatorial optimization, parallel processing, image processing, geographic information systems, data base systems, and data structures. In many applications it is necessary to number the points of a discrete multi-dimensional space (or, equivalently, a grid) by an indexing scheme mapping each point bijectively to a natural number in the range between 1 and the total number of points in the space. Often it is desirable that this indexing scheme preserves some kind of locality, that is, close-by points in the space are mapped to close-by numbers or vice versa. For this purpose, indexing schemes based on space- lling curves have shown to be of high value [4{9]. In this paper we study Hilbert indexings, perhaps the most popular space lling indexing schemes. Properties of 2D and 3D Hilbert indexings have been extensively studied recently [4{10]. However, most of the work so far has focused on empirical studies. Up to now, little attention has been paid to the theoretical study of structural properties of multi-dimensional Hilbert curves, the focus of this paper. Whereas with \modulo symmetry" there is only one 2D Hilbert curve, there are many possibilities to de ne Hilbert curves in the 3D setting [4, 9]. The advantage of Hilbert curves is their (compared to other curves) simple structure. Our results can shortly be sketched as follows. We generalize the notion of Hilbert indexings to arbitrary dimensions. We clarify the concept of Hilbert curves in multi-dimensional spaces by providing a natural and simple mathematical formalism that allows combinatorial studies of multi-dimensional Hilbert ?

Work partially supported by a Feodor Lynen fellowship of the Alexander von Humboldt-Stiftung, Bonn, and the Center for Discrete Mathematics, Theoretical Computer Science, and Applications (DIMATIA), Prague.


indexings. For reasons of (geometrical) clearness, we base our formalism on permutations instead of e.g. matrices or other formalisms [2{4, 10]. So we obtain the following insight: Space- lling curves with Hilbert property can be completely described by simple generating elements and permutations operating on them. Structural questions for Hilbert curves in arbitrary dimensions can be decided by reducing them to basic generating elements. Putting it in catchy terms, one might say that for Hilbert indexings what holds \in the large" (i.e., for large side-length), can already be detected \in the small" (i.e., for side-length 2). In particular, this provides a basis for mechanized proofs of locality of curves with Hilbert property (cf. [9]). In addition, this observation allows the identi cation of seemingly dierent 3D Hilbert indexings [4], the generalization of a locality result of Gotsman and Lindenbaum [6] to a larger class of multi-dimensional indexing schemes, and the determination that there are exactly 6 28 = 1536 structurally dierent 3D Hilbert curves. The latter clearly generalizes and answers Sagan's quest for describing 3D Hilbert curves [10]. Finally, we provide an easy recursive formula for computing Hilbert indexings in arbitrary dimensions and sketch a recipe for how to construct an r-dimensional Hilbert curve for arbitrary r in an easy way from two (r ? 1)-dimensional ones. Some missing details and proofs can be found in the full version of the paper [1].

2 Preliminaries We focus our attention on cubic grids, where, grid of side-length n. An rdimensional (discrete) curve C is simply a bijective mapping C : f1; : : : nr g ! f1; : : : ; ngr . Note that, by de nition, we do not claim the continuity of a curve. A curve C is called continuous if it forms a Hamilton path through the nr grid points. An r-dimensional cubic grid is said to be of order k if it has side-length 2k . Analogously a curve C has order k if its range is a cubic grid of order k. 2

3

11111111 00000000 00000000 11111111

1

4

1 0

4

1

11 00

1 0 1 0

3

11 00 00 11

2

11111111 00000000 00000000 11111111 11 0 00 1

Fig. 1. The generator Hil21 and its

g. canonical corner-indexing Hil 2 1

111 000

Fig. 2. Construction scheme for the 2D Hilbert indexing.

Fig. 1 shows the smallest 2D continuous curve indexing a grid of size 4. This curve can be found in Hilbert's original work (see [11]) as a constructing unit for a whole family of curves. Fig. 2 shows the general construction principle for these so-called Hilbert curves: For any k 1 four Hilbert indexings of size 4k


are combined into an indexing of size 4k+1 by rotating and re ecting them in such a way that concatenating the indexings yields a Hamilton path through the grid. One of the main features of the Hilbert curve is its \self-similarity". Here \self-similar" shall simply mean that the curve can be generated by putting together identical (basic construction) units, only applying rotation and re ection to these units. In a sense, the Hilbert curve is the \simplest" self-similar, recursive, locality-preserving indexing scheme for square meshes of size 2k 2k .

3 Formalizing Hilbert curves in r dimensions In this section, we generalize the construction principle of 2D Hilbert curves to arbitrary dimensions in a rigorous, mathematically precise way.

3.1 Classes of Self-Similar Curves and their generators Let Vr := fx1 x2 xr?1 xr j xi 2 f0; 1g g be the set of all 2r corners of an r-dimensional cube coded in binary. Moreover, let I : Vr ?! f1; : : : ; 2r g denote

an arbitrary indexing of these corners. To describe the orientation of subcurves inside a curve of higher order, we want to use symmetry mappings, which can be expressed via suitable permutations operating on such corner-indexings. Observe that any r-dimensional curve C1 of order 1 naturally induces an indexing of these corners (see Fig. 1 and Fig. 3). We call the obtained corner-indexing the canonical one and denote it by Cf1 : Vr ?! f1; : : : ; 2r g. Furthermore, let WI denote the group of all permutations (operating on I ) that describe rotations and re ections of the r-dimensional cube. In other words, WI is the set of all permutations that preserve the neighborhood-relations n(i; j ) of the corner indexing I :

WI := 2 Sym(2r ) : n(i; j ) = n((i); (j )) 8i; j 2 f1; : : : ; 2r g : For a given permutation 2 WI , we sometimes write ( : I ) in order to emphasize that is operating on a cube with corner-indexing I . The point here is that once we have xed a certain corner-indexing I , the set WI will provide all

necessary transformations to describe a construction principle of how to generate curves of higher order by piecing together a suitable curve of lower order. Obviously each permutation ( : I ) acting on a given corner-indexing I canonically induces a bijective mapping on a cubic grid of order k. Subsequently, we do not distinguish between a permutation and the corresponding mapping on a grid. We partition an r-dimensional cubic grid of order k into 2r subcubes of order k ? 1. For each x1 xr 2 Vr we therefore set

p((kx) xr ) := (x1 2k?1 ; : : : ; xr 2k?1 ) 2 f0; : : : ; 2k ? 1g : : : f0; : : : ; 2k ? 1g 1

to be the \lower-left corner" of such a subcube. Let Ck?1 be an r-dimensional curve of order k ? 1 (k 2). Our goal is to de ne a \self-similar" curve Ck of order k by putting together 2r pieces of type Ck?1 . Let I : Vr ?! f1; : : : ; 2r g be a corner-indexing. We intend to arrange the 2r subcurves of type Ck?1 \along"


I . The position of the i0 -th (where i0 2 f1; : : : ; 2r g) subcurve inside Ck can for-

mally be described with the help of the grid-points p((kx) xr ) . Bearing in mind the classical construction principle for the 2D Hilbert indexing, the orientation of the constructing curve Ck?1 inside Ck can be expressed by using symmetric transformations (that is re ections and rotations). For any sequence of permutations 1 ; : : : ; 2r 2 WI we therefore de ne 1

?

(1) Ck (i) := (i0 : I ) Ck?1 i mod (2k?1 )r + pI(k?) (i0 ) ; where i 2 f1; : : : ; (2k )r g and i0 = (i ? 1) div (2k?1 )r + 1. The geometric intuition behind is that the curve Ck can be partitioned into 2r components of the form Ck?1 (re ected or rotated in a suitable way). These subcurves are arranged inside Ck \along" the given corner-indexing I . The orientation of the i0 -th subcurve inside Ck is described by the eect of i0 operating on I . De nition 1. Whenever two r-dimensional curves Ck?1 of order k ? 1 and Ck of order k satisfy equation (1) for a given sequence of permutations 1 ; : : : ; 2r 2 WI (operating on the corner-indexing I : Vr ?! f1; : : : ; 2r g), we will write Ck?1 ( ;::: ; Ir ) Ck and call Ck?1 the constructor of Ck . Our nal goal is to iterate this process starting with a curve C1 of order 1. 1

1

2

It's only natural and in our opinion \preserves the spirit of Hilbert" to x the corner-indexing according to the structure of the de ning curve C1 . Hence, in this situation we can specify our I to be the canonical corner-indexing Cf1 . By successively repeating the construction principle in equation (1) k times, we obtain a curve of order k.

De nition 2. Let C = f Ck j k 1 g be a family of r-dimensional curves of order k. We call C a Class of Self-Similar Curves (CSSC) if there exists a sequence of permutations 1 ; : : : ; 2r 2 WCf (operating on the canonical cornerindexing Cf1 ) such that for each curve Ck it holds that 1

C1

f

(1 ;::: ;2r )

C2

C1

f

(1 ;::: ;2r ) C1

f

(1 ;::: ;2r ) C1

Ck?1

f

(1 ;::: ;2r ) C1

Ck :

In?this case, C1 is called the generator of the CSSC C and we de ne the set H C1 ; (1 ; : : : ; 2r ) := f Ck j k 1 g to be the CSSC generated by C1 and 1 ; : : : ; 2r . A CSSC C = f Ck j k 1 g is called Class with Hilbert Property

(CHP) if all curves Ck are continuous. ?

Note that the CSSC H C1 ; (1 ; : : : ; 2r ) is well-de ned, because any CSSC is uniquely determined by its generator C1 and the choice of the permutations 1 ; : : : ; 2r 2 WCf . Our concept for multi-dimensional CHPs only makes use of the very essential tools which can be found in Hilbert's context (cf. [11]) as rotation and re ection. We deliberately avoid more complicated structures (e.g., the use of dierent sequences of permutations in each inductive step, or the use of several generators for the constructing principle) in order to maintain conceptual 1


simplicity and ease of construction and analysis. However, the theory which we develop in this paper doesn't necessarily restrict to the continuous case. We end this subsection with an example. ?One easily checks that the classical 2D Hilbert indexing can be described via H Hil21 ; ((2 4); id; id; (1 3)) = f Hil2k j k 1 g, where the generator Hil21 is given in Fig. 1. As Theorem 2 will show, this is the only CHP of dimension 2 \modulo symmetry."

3.2 Disturbing the generator of a CSSC In this subsection we analyze the eects of disturbing the generator of a CSSC by a symmetric mapping. We will see that any disturbance of the generator will be hereditary to the whole CSSC in a very canonical way. And also the other way round: if two dierent CSSCs show a certain similarity in one of their members, this similarity can already be found in the structure of the corresponding? generators. We illustrate this by the following diagram. Given two CSSCs ? H C1 ; (1 ; : : : ; 2r ) = fCk j k 1g and H D1 ; (1 ; : : : ; 2r ) = fDk j k 1g, respectively.1 Suppose there is a similarity at a certain stage of the construction, i.e., for some k0 the curves Ck and Dk can be obtained from each other by a similarity transformation . The investigations in this section will show that the inner structure of CSSCs are strong enough to yield the same behavior at the stage of any order. 0

0

C?1 ? ? ? y

D1

f

1 (1 ;::: ;2r ) C?2

C

?

? ? y f D ( ;::: ; r ) D2 1

1

2

f

(1 ;::: ;2r ) ? C1

?

? ? y f D ( ;::: ; r ) 1

1

2

f

(1 ;::: ;2r ) C?k C1

?

? ? y f D ( ;::: ; r ) Dk 1

1

2

Consequently, for issues like structural behavior, it will be sucient to analyze the generating elements of a CSSC only, since we nd all necessary information encoded here. We start with a simple observation concerning the behavior of the construction principle of De nition 1 under the \symmetric disturbance" of a constructor. We omit the proof. Lemma 1. Let Ck?1 and Ck be curves of order k ?1 and k, respectively. Suppose Ck?1 is the constructor of Ck , i.e., Ck?1 ( ;::: ; Ir ) Ck , for any sequence of permutations 1 ; : : : ; 2r 2 WI (acting on a given corner-indexing I ). Then for arbitrary 2 WI we have 1

( : I ) Ck?1

2

I (1 ?1 ;::: ;2r ?1 )

Ck :

Whereas, by Lemma 1, we investigated the in uence of disturbing the constructor, we now, in a second step, analyze how transforming the underlying 1

Note that the 's used in the de nition of both CSSCs yield completely dierent automorphisms on the grid. Whereas in the rst case they refer to the corner-indexing f1 , given by generator D1 . Cf1 , in the second case they act on the corner-indexing D


corner-indexing in uences the construction principle. We will need such a result, since two dierent CSSCs (by de nition) come up with two dierent cornerindexings, each of which given by the underlying generator. Again we omit the proof. Lemma 2. Given the assumptions of Lemma 1 (that is: Ck?1 ( ;::: ; rI) Ck for two curves Ck?1 and Ck of successive order), then for arbitrary 2 WI and the modi ed corner-indexing K := ?1 I with = ( : I ) = ( : K) we have2 1

Ck?1

K

(1 ;::: ;2r )

2

Ck :

Lemma 1 and 2 now allow the proof of the main result of this section. For its illustration we refer to the diagram at the beginning of this section. Do also recall the point made in the footnote there. ? Theorem 1. Let C1 be the generator of the CSSC H C1 ; (1 ; : : : ; 2r ) = fCk j ? k 1g and D1 the generator of the CSSC H D1 ; (1 ; : : : ; 2r ) = fDk j k 1g. For an arbitrary permutation 2 WCf and the corresponding symmetric f1 ), the following statements are equivalent: mapping = ( : Cf1 ) = ( : D (i) Ck = Dk for some k0 1. (ii) Ck = Dk for all k 1. Proof. (ii) ) (i) is trivial. For (i) ) (ii) we rst show that statement (ii) is true for the generators C1 and D1 : If k0 > 1 we can divide the cubic grid of order k0 into 2r subgrids of order k0 ? 1. By the construction principle for CSSCs, the curves Ck and Dk traverse these subgrids \along" the canonical cornerf1 . Since, by assumption, Ck = Dk , the corresponding indexings Cf1 resp. D relation also holds true for the corner-indexings Cf1 and Df1 , which nally yields the validity of the equation C1 = D1 , because of the isomorphisms C1 ' Cf1 f1 . We proceed proving (ii) by induction on k . Assuming that resp. D1 ' D Dk = Ck we show this relation for k + 1 by applying Lemma 1 and Lemma 2. Since fCk j k 1g is a CSSC, we get =) 1 | {zCk} ( ? ;::: ; rCf Ck ( ;::: ;Cfr ) Ck+1 Lemma ? ) Ck+1 1

0

0

0

0

0

0

1

1

2

Lemma 2

=Dk

1

1

1

2

1

f

1 (1 ;::: ;2r ) Ck+1 ; where the last relation makes use of Df1 = ?1 Cf1 , which we immediately obtain from the given equation D1 = C1 .3 This implies Dk+1 = Ck+1 because

=) Dk

D

of the CSSC-property of fDk j k 1g. ut In particular, the result of Theorem 1 implies that any questions concerning the structural similarity of two CSSCs can be reduced to the analysis of their generators. The fact that the corner-indexing is disturbed by ?1 instead of is due to technical reasons only. 3 A disturbance by implies a transformation of the corner-indexings by ?1 , which can be easily checked.

2


3 4

6

3

8

6

5 2

5

4

2

1

generator Hil31 .A

5

4

3

6 8

7

6 1

7 8

1

generator Hil31 .B

8

8

7 2

3

2

7

2 1

4

5

7

3

7

8 3

6

5

6 1

2

4

1

4

5

generator Hil31 .C

g31.x. Fig. 3. Continuous 3D generators Hil31 .x and their canonical corner-indexings Hil

4 Applications: Computing and analyzing CHPs First in this section, we attack a classi cation of all structurally dierent CHPs for higher dimensions. Whereas we can provide concrete combinatorial results for the 2D and 3D cases, the high-dimensional cases appear to be much more dicult. The basic tool for such an analysis, however, is given by Theorem 1. The following theorem justi es the naming \class with Hilbert property" (CHP). Theorem 2. The classical 2D Hilbert indexing H? Hil21 ; ((2 4); id; id; (1 3)) is the only CHP of dimension 2 modulo symmetry. Proof. Due to Theorem 1 it suces to show that Hil21 is the only continuous 2D generator, which is obvious. In addition, we have to check whether there is another sequence of permutations such that 4 generators Hil21 can be arranged in g2 a grid of order 2 along the canonical corner-indexing Hil 1 in a continuous way. A simple combinatorial consideration shows that no other sequence of permutations yields a continuous curve of order 2 whose starting- and endpoints are located at corners of the grid. However, any constructor for a continuous curve of higher order must have this property. ut What about the 3D case? The analysis of the \Simple Indexing Schemes" (which are related to our CHPs) in Chochia and Cole [4] already shows that the number of CHPs in the 3D case grows drastically compared to the 2D setting. However, by our analysis, lots of \Simple Indexing Schemes" in [4] now turn out to be identical modulo symmetry. We state the following classi cationtheorem, which treats the 3D case entirely. It also generalizes and answers work of Sagan [10]. Theorem 3. For the 3D case there are 6 28 = 1536 structurally dierent (that is: not identical modulo re ection and rotation) CHPs. These types are listed in Table 1. Proof (Sketch). Theorem 1 says that we can restrict our attention to checking any continuous curves of order 1 which are dierent modulo symmetry. Given such a continuous generator C , the total amount of CHPs which can be constructed by C is given by all possibilities of piecing together 8 (rotated or re ected) versions of C (\subcurves") along its canonical corner-indexing Ce. By exhaustive search, we get that there are 3 dierent (modulo symmetry) types


Table 1. Description of all 3-dimensional CHPS. generator version 1 2 3 4 (2 8)(3 5) / (3 7)(4 8) / (3 7)(4 8) / (1 3)(6 8) / (a) (2 4 8)(3 5 7) (2 8 4)(3 7 5) (2 8 4)(3 7 5) (1 3)(2 4)(5 7)(6 8) (2 8)(3 5) / (3 7)(4 8) / id / (1 7 3)(4 6 8) / (b) (2 4 8)(3 5 7) (2 8 4)(3 7 5) (2 4)(5 7) (1 7 5 3)(2 8 6 4) Hil31 .A (2 8)(3 5) / (3 7)(4 8) / id / (1 7)(4 6) / (c) (2 4 8)(3 5 7) (2 8 4)(3 7 5) (2 4)(5 7) (1 7 5)(2 6 4) (2 8)(3 5) / (3 7)(4 8) / (3 7)(4 8) / (1 3)(6 8) / (d) (2 4 8)(3 5 7) (2 8 4)(3 7 5) (2 8 4)(3 7 5) (1 3)(2 4)(5 7)(6 8) (2 8)(5 7) / id / (3 5)(6 8) / (2 8)(5 7) / (a) (2 6 8)(3 5 7) (2 6)(3 7) (2 8 6)(3 7 5) (2 6 8)(3 5 7) Hil31 .B (2 8)(5 7) / id / (3 5)(6 8) / (3 5)(6 8) / (b) (2 6 8)(3 5 7) (2 6)(3 7) (2 8 6)(3 7 5) (2 8 6)(3 7 5) generator version 5 6 7 8 (1 3)(6 8) / (1 5)(2 6) / (1 5)(2 6) / (1 7)(4 6) / (a) (1 3)(2 4)(5 7)(6 8) (1 5 7)(2 4 6) (1 5 7)(2 4 6) (1 7 5)(2 6 4) (1 3 5)(2 6 8) / id / (1 5)(2 6) / (1 7)(4 6) / (b) (1 3 5 7)(2 4 6 8) (2 4)(5 7) (1 5 7)(2 4 6) (1 7 5)(2 6 4) Hil31 .A (2 8)(3 5) / id / (1 5)(2 6) / (1 7)(4 6) / (c) (2 4 8)(3 5 7) (2 4)(5 7) (1 5 7)(2 4 6) (1 7 5)(2 6 4) 6 8) / id / (1 5)(2 6) / (1 7)(4 6) / (d) (1(13355)(2 7)(2 4 6 8) (2 4)(5 7) (1 5 7)(2 4 6) (1 7 5)(2 6 4) (1 3)(4 6) / (1 3)(4 6) / id / (1 7)(2 4) / (a) (1 3 7)(2 4 6) (1 3 7)(2 4 6) (2 6)(3 7) (1 7 3)(2 6 4) Hil31 .B (1 7)(2 4) / (1 3)(4 6) / id / (1 7)(2 4) / (b) (1 7 3)(2 6 4) (1 3 7)(2 4 6) (2 6)(3 7) (1 7 3)(2 6 4)

of continuous generators, namely Hil31 .A, Hil31 .B and Hil31 .C (see Fig. 3). As described above, we now have to check whether there are continuous arrangements of these generators along their canonical corner-indexings. Beginning with type A, an exhaustive combinatorial search yields that there are 4 possible cong3 tinuous formations of Hil31 .A along Hil 1 .A. All possibilities are shown in Fig. 4, where the orientation of each subcube is given by the position of an edge (drawn in bold lines). For each subcube there are two symmetry mappings which yield possible arrangements for the generator within such a subgrid. The permutations expressing these mappings are listed in Table 1. Analogously, we nd out the possible arrangements for generator type B. Note that there are no more than 2 dierent continuous arrangements of this generator along its canonical corner-indexing. Finally we easily check that Hil31 .C cannot even be the constructor of a continuous curve of order 2. Table 1 thus yields that there are exactly 4 28 + 2 28 = 6 28 structurally dierent CHPs. ut Construction of an r-dimensional Hilbert curve. Without giving an explicit proof here, we just indicate how the construction of a high-dimensional CHP can be done inductively: A continuous generator of dimension r can be derived simply by \joining together" two continuous generators of dimension r ? 1. As an example we give a CHP of dimension 4, whose generator Hil41 is constructed by joining together two generators Hil31 , version (a) (cf. Figure 3). The generator Hil41 and a suitable sequence of permutations are shown in Fig. 5. Note that this construction principle can be extended to obtain Hilbert indexings in arbitrary dimensions in an expressive and easy, constructive way: Following the


Hil31 .A-version (a)

Hil31 .A-version (b)

Hil31 .A-version (c)

Hil31 .A-version (d)

Hil31 .B-version (a)

Hil31 .B-version (b)

Fig. 4. Construction principles for CHPs with generators Hil31.A and Hil31 .B. construction principle of Hil31 , version (a), rst pass through an r ? 1-dimensional structure, then in \two steps" do a change of dimension in the rth dimension, and nally again pass through an r ? 1-dimensional structure. This method applies to nding the generators as well as to nding the permutations.

Recursive computation of CSSCs. Note that whenever a CSSC C = fCk j k 1g is explicitly given by its generator and the sequence of permutations, we

may use the recursive formula (1) of Subsection 3.1 to compute the curves Ck . In other words, the de ning formula (1) itself provides a computation-scheme for CSSC, which is parameterized by the generating elements (generator and sequence of permutations).

Aspects of locality. The above mentioned parameterized formula might, for

example, also be used to investigate locality properties of CSSCs by mechanical methods. The locality properties of Hilbert curves have already been studied in great detail. As an example for such investigations, we brie y note a result of Gotsman and Lindenbaum [6] for multi-dimensional Hilbert curves. In [6] they investigate a curve C : f1; : : : ; nr g ! f1; : : : ; ngr with the help of their locality measure L2 (C ) = maxi;j2f1;::: ;nr g (d2 (C (i); C (j )))=ji ? j j, where d2 denotes the Euclidean metric. In their Theorem 3 they claim the upper bound L1(Hkr ) r r (r + 3) 2 for any r-dimensional Hilbert curve of order k, without precisely specifying what an r-dimensional Hilbert curve shall be. Since the proof of their result does not utilize the special Hilbert structure of the curve, this result can even be extended to arbitrary CSSCs. 2


3

6

4

5 11

14 13

12 15

16 2 1

10 9

7 8

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

= = = = = = = = = = = = = = = =

(2 16)(3 13)(6 12)(7 9) (3 15)(4 16)(5 9)(6 10) 2 (1 3 13 11 9 7)(2 4 14 12 10 8)(5 15)(6 16) 4 (1 5 13 9)(2 6 14 10)(3 11 15 7)(4 12 16 8) 6 (1 7)(4 6)(10 16)(11 13) 8 (1 9 13 5)(2 10 14 6)(3 7 15 11)(4 8 16 12) 10 (1 11)(2 12)(3 5 7 9 15 13)(4 6 8 10 16 14) 12 (1 13)(2 14)(7 11)(8 12) 14 (1 15)(4 14)(5 11)(8 10)

Fig. 5. Constructing elements for a 4-D CHP (generator Hil41 and permutations).

5 Conclusion Our paper lays the basis for several further research directions. So it could be tempting to determine the number of structurally dierent r-dimensional curves with Hilbert property for r > 3. Moreover, a (mechanized) analysis of locality properties of r-dimensional (r > 3) Hilbert curves is still to be done (cf. [9]). An analysis of the construction of more complicated curves using more generators or dierent permutations for dierent levels remains open.

References

1. J. Alber and R. Niedermeier. On multi-dimensional Hilbert indexings. Technical Report 98-392, KAM-DIMATIA Series, Faculty of Mathematics and Physics, Charles University, Praha, Czech Republic, May 1998. 2. A. R. Butz. Space lling curves and mathematical programming. Information and Control, 12:314{330, 1968. 3. A. R. Butz. Convergence with Hilbert's space lling curve. Journal of Computer and System Sciences, 3:128{146, 1969. 4. G. Chochia and M. Cole. Recursive 3D mesh indexing with improved locality. Technical report, University of Edinburgh, 1997. Short version appeared in the Proceedings of HPCN'97, LNCS 1225. 5. G. Chochia, M. Cole, and T. Heywood. Implementing the Hierarchical PRAM on the 2D mesh: Analyses and experiments. In Symposium on Parallel and Distributed Processing, pages 587{595. IEEE Computer Science Press, October 1995. 6. C. Gotsman and M. Lindenbaum. On the metric properties of discrete space- lling curves. IEEE Transactions on Image Processing, 5(5):794{797, May 1996. 7. H. V. Jagadish. Linear clustering of objects with multiple attributes. ACM SIGMOD Record, 19(2):332{342, June 1990. 8. I. Kamel and C. Faloutsos. Hilbert R-tree: An improved R-tree using fractals. In 20th International Conference on Very Large Data Bases, pages 500{509, 1994. 9. R. Niedermeier, K. Reinhardt, and P. Sanders. Towards optimal locality in meshindexings. In Proceedings of the 11th Conference on Fundamentals of Computation Theory, number 1279 in Lecture Notes in Computer Science, pages 364{375, Krakow, Poland, Sept. 1997. Springer-Verlag. 10. H. Sagan. A three-dimensional Hilbert curve. Int. J. Math. Educ. Sci. Technol., 24(4):541{545, 1993. 11. H. Sagan. Space-Filling Curves. Universitext. Springer-Verlag, 1994.